The problem
Agent runs end early: budgets run out, sessions end, workers get cut off mid-task. Each ending takes the run's progress with it, and the next attempt starts from zero, guessing what the last one finished.
The build
- Built a checkpoint plugin that writes the run's state to the record as it works: what it finished, what comes next, and the evidence so far.
- Checkpoints land at progress tiers through the run and once more before the budget closes, so the hand-off to the next attempt is a document, not a guess.
- Ran it across my own production fleet first, where cut-off runs were a weekly cost.
The evidence
- Published open source under MIT at v0.3.0 (October 2026), released alongside the hermes-muse-bridge and muse-pstack projects.
- In production on the fleet that operates my businesses: a run that hits its cap now hands its progress to the next attempt.
Long agent runs deserve the same courtesy as long build jobs: a record of where they stopped. agent-checkpoint is the plugin I built for that. It persists a run’s progress as the run works, so time and budget limits end a run cleanly, with its work intact and its next steps written down. It came out of my own fleet’s operations, and it runs there today.